Less Risk, More Reward: Proving Your AI Strategy Works

When it comes to AI, many organisations are trying to translate aspirational claims into achievable outcomes. What can it really do, and what can it really do for me?

To help answer those questions, we’re going to look at the three stages every organisation should go through to mitigate risk and maximise reward. These are:

• Business justification
• Success criteria
• Implementation

These should be the core of any business decision, but with AI tools, the potential risks and rewards are particularly stark. This makes it all the more important you get it right.

Before we go any further, though, let’s set out a suitable example:

A fee-earner business with 100 employees. Increased regulatory overhead, a growing data estate, and changing client demands mean each job is taking longer to complete.

This is driving client costs up in an already competitive environment, as well as hitting the profitability of fixed-fee engagements. Business is being lost due to client cost concerns and margins are being eroded. With every new acquisition or regulation, further complexity means these problems only get worse.

With that reference point, let’s deconstruct the process.

 

Step 1: Business Justification

Start by inverting your problem statement. It needn’t even be framed as a negative – wanting a competitive advantage is a problem, if you don’t currently have one. Building a strength can be just as powerful a motivator as mitigating a weakness.

In our example, it’s very simple. We’ve already identified being uncompetitive and less profitable due to a high-cost base as the starting point. The destination is an improved competitive position which results in more business being won, at better margins.

This can be achieved by reducing client costs and controlling delivery times for fixed-fee work. Inefficiencies and rising workloads are driving costs up and profits down, so increasing productivity is a clear path to addressing both these issues. Diagnosis complete, step one done?

Not quite. It’s around this stage that things tend to wobble.

In our example, that’s because “more productivity” is a slogan, not a plan. It’s a statement of aspiration to a general concept. Nobody has ever aimed to be less productive. What does valuable productivity gain look like for you?

For some, productivity means less time taken revising client reports. For others, it means providing relevant, context-aware information without just turning time spent searching into time spent validating. In all cases, it means understanding your workflows, people, and what’s of value to you.

Failure to define that is a common cause of AI strategies falling apart at the first hurdle.

In this situation, hope acts as a substitute for evidence. Licences get purchased and nobody knows whether there is a return associated with that investment.

Measurement requires definition, which we look at in step two .

 

Step 2: Success Criteria

Having confidence that a plan should work and being sure it has worked are very different things. With only 18% of professional business services organisations tracking the return on their AI investment, this is a big gap.

Addressing this is step two. But why do we place it before implementation steps – isn’t that counterintuitive?

It’s to avoid a framing trap. If you start with a plan, the success criteria will shift to become “did we deliver the plan?” rather than “did the plan deliver the desired outcomes?” Meeting deadlines and ticking feature boxes is a successful project. It is not a successful strategy.

By establishing the success criteria prior to the implementation plan, we make sure that we’re measuring progress in the context of strategic business objectives. This ensures measurability is designed into the implementation plan, not fitted around it in retrospect.

Many organisations are struggling with this point in relation to AI. Outcomes should be aspirational, but they shouldn’t be only aspirations. When they’re vague, they become assumed outcomes of following the plan. If you do X, then Y somehow happens. That’s the business equivalent of a cargo cult.

Conversely, when they’re defined, they become valuable metrics which provide further business intelligence. They can guide where you should make further investment or flag the advent of diminishing returns. They inform which AI tool you should choose – and it’s at the end of step 2 that you should make that call.

What these criteria are will vary between organisations. There will be commonalities, but context matters a lot and will shape what good looks like for you. However, amongst the many success criteria from which you can choose, there is one failure condition, and it is universal.

You have to do all this safely and the ICO has made it clear the responsibility for this cannot simply be delegated to technical teams.

 

Step 3: Implementation

This falls into two phases: preparing the ground and then building on it.

Preparing the ground is the more complex of the two. It requires deep expertise in things such as data security or architecture, often with further specialisation into particular vendor technologies.

It can also be much broader and deeper than just data concerns.

Dawson Cornwell’s Azure migration built a foundation that was AI-ready, rather than retrofitting AI to systems predating its existence. The infrastructure transformation came first. This allowed AI to be adopted safely and securely. Data stores were restructured, labels applied, and guardrails established through policy-driven governance tools.

It’s a deliberately extreme example – the groundwork isn’t always so involved. What you cannot avoid, even in the most modest of scenarios, is security and compliance.

The key principles here are understanding your data, putting guardrails around it, and being able to demonstrate their efficacy while also delivering your business objectives. It means not deploying an AI tool into an ungoverned Microsoft 365 environment. It means knowing what sensitive data you have, where it is, and being able to demonstrate how you’re keeping it safe.

In such scenarios, regulatory bodies would tend towards more severe penalties, should a breach occur. This is why you need to have a demonstrable, auditable process before you finally move on to the building phase.

Once you’re ready to enable the AI tools themselves, you need to provide adoption training – first to your champions, then later to the wider rings of users. This systematic, standardised onboarding accelerates benefits and makes your continuous improvement process diagnostic, rather than relying on trial and error.

Rather than throwing money at AI in the hope it solves more problems than it causes, this approach establishes a predictable, strategically aligned methodology with repeatable proof at its core.

Our example business has not only achieved its stated objectives but also mitigated substantial risks – financial and reputational – and reduced its future regulatory reporting overhead. All this while gaining a competitive advantage, built on strong foundations and with proven success.